Streaming & the work log
Follow live output, stop generation, and read safe progress summaries.
Responses stream into the chat pane as they are generated. Convography also preserves a safe work log when the provider or route supplies progress steps or a reasoning summary.
Streaming states#
After you send, the graph adds a pending node and the chat pane shows an active response. Text appears incrementally rather than waiting for the complete answer. A caret marks the live edge of the response.
When generation finishes, Convography stores the final content, model metadata, citations, attachments, usage details, and completion status. Reloading the conversation shows the persisted answer rather than replaying the stream.
Read the work log#
The work log is labeled Thinking and can include steps marked active, done, or failed plus a concise reasoning summary. It is useful for seeing whether a request is researching, processing files, generating an image, or preparing an answer.
The log is not a transcript of private model reasoning. It contains only safe, supported summaries and progress events. Some routes provide no log, in which case the pane shows a normal loading state until answer text arrives.
Stop a response#
Select Stop to abort the active browser request. The pending graph state is cleaned up and you can revise the prompt. Depending on timing, upstream work and associated usage may already have occurred.
Navigate while work continues#
You can select another node or conversation while a response streams. Convography tracks each active generation by its original tree and parent. When a background response completes, it remains attached to the correct branch. It does not replace your newer selection.
Failed responses#
A provider or persistence error can produce a failed response state. Successful sibling results from Send all remain available. Open the failed node to see the user-facing error, then decide whether to retry, change route, reduce file context, or wait for provider recovery.
Avoid sending the same request many times during an outage. Each accepted attempt can consume usage even if the final experience is unsuccessful.